model = Model(2048, 101, np.random.randint(20, size=20))
data = torch.randn((1, 750, 2048))
macs, params = profile(model, inputs=(data, 'cpu', False,torch.from_numpy(
np.array([138, 391, 286, 206, 102, 750, 385, 193, 198, 325, 318, 347, 280, 700,
228, 202, 147, 254, 283, 185, 219, 321, 310, 583, 62, 226, 175, 750,
316, 255, 89, 51])) ))
FLOPs = macs * 2
print('FLOPs: %d' % FLOPs) # 6601728000
print('params: %d' % params) # 4405450
from facebookresearch_params_flops_remove_g import params_count, flop_count
print('facebookresearch params: %d' % params_count(model)) # 4405652
flop_dict, _ = flop_count(model, (data,)) # 4401152
flops = sum(flop_dict.values())
print('facebookresearch flops: %d' % flops)
model = Model(2048, 101, np.random.randint(20, size=20)).cuda()
data = torch.randn((1, 750, 2048)).cuda()
t1 = time.time()
for i in range(1000):
res = model(data, torch.device("cuda"), is_training=False, seq_len=torch.from_numpy(
np.array([138, 391, 286, 206, 102, 750, 385, 193, 198, 325, 318, 347, 280, 700,
228, 202, 147, 254, 283, 185, 219, 321, 310, 583, 62, 226, 175, 750,
316, 255, 89, 51])).cuda())
if __name__ == '__main__':
import sys
sys.path.append('/disk3/zt/code/2_TIP_rebuttal/5_FLOPs_inference_time/lib')
from config import cfg, update_config
cfg_file = '/disk3/zt/code/2_TIP_rebuttal/5_FLOPs_inference_time/experiments/thumos/A2Net.yaml'
update_config(cfg_file)
model = GTAN(cfg)
data = torch.randn((1, 2048, 512))
macs, params = profile(model, inputs=(data, ))
FLOPs = macs * 2
print('FLOPs: %d' % FLOPs) # 5011222384
print('params: %d' % params) # 107635264
from facebookresearch_params_flops_remove_g import params_count, flop_count
print('facebookresearch params: %d' % params_count(model)) # 107635264
flop_dict, _ = flop_count(model, (data, )) # 2504884224
flops = sum(flop_dict.values())
print('facebookresearch flops: %d' % flops)
# Inference Time
import time
model = GTAN(cfg).cuda()
data = torch.randn((1, 2048, 512)).cuda()
t1 = time.time()
for i in range(1000):
res = model(data)
t2 = time.time()
print("Inference Time:", (t2 - t1) / 1000) # 0.01113538670539856
# """
# return np.sum([p.numel() for p in model.parameters()]).item()
#
#
# print('facebookresearch params: %d' % params_count(model))
#
# from fvcore.nn.flop_count import flop_count
#
# gflop_dict, _ = flop_count(model, (data,)) #36978564
# gflops = sum(gflop_dict.values())
# print('facebookresearch gflops: %d' % gflops) # 91
from facebookresearch_params_flops_remove_g import params_count, flop_count
print('facebookresearch params: %d' % params_count(model)) # 36978564
flop_dict, _ = flop_count(model, (data, )) # 91248435200
flops = sum(flop_dict.values())
print('facebookresearch flops: %d' % flops)
model = BMN(opt).cuda()
input = torch.randn((1, 400, 100)).cuda()
t1 = time.time()
for i in range(1000):
a, b, c = model(input)
t2 = time.time()
print((t2 - t1) / 1000)
# 0.016720424890518188
# 0.020179431200027464
# 0.020179431200027464
# 0.02353054356575012
# model (model): model to count the number of parameters.
# """
# return np.sum([p.numel() for p in model.parameters()]).item()
#
#
# print('facebookresearch params: %d' % params_count(model))
#
# from fvcore.nn.flop_count import flop_count
#
# gflop_dict, _ = flop_count(model, (data,)) #65527534
# gflops = sum(gflop_dict.values())
# print('facebookresearch gflops: %d' % gflops) # 1
from facebookresearch_params_flops_remove_g import params_count, flop_count
print('facebookresearch params: %d' % params_count(model)) #65527534
flop_dict, _ = flop_count(model, (data, )) # 1239018496
flops = sum(flop_dict.values())
print('facebookresearch flops: %d' % flops)
# Inference Time
import time
model = LocNet(cfg).cuda()
data = torch.randn((1, 2048, 128)).cuda()
t1 = time.time()
for i in range(1000):
res = model(data)
t2 = time.time()
print("Inference Time:",
(t2 - t1) / 1000) # 0.013252334356307984 small change every time
cfg_file = '/disk3/zt/code/2_TIP_rebuttal/5_FLOPs_inference_time/experiments/thumos/A2Net.yaml'
update_config(cfg_file)
model = SSAD(cfg)
data = torch.randn((1, 303, 512))
# out = model(data)
# for i in out:
# print(i.size())
macs, params = profile(model, inputs=(data, ))
FLOPs = macs * 2
print('FLOPs: %d' % FLOPs) # 905643072
print('params: %d' % params) # 3735608
from facebookresearch_params_flops_remove_g import params_count, flop_count
print('facebookresearch params: %d' % params_count(model)) # 3735608
flop_dict, _ = flop_count(model, (data, )) # 452640768
flops = sum(flop_dict.values())
print('facebookresearch flops: %d' % flops)
# Inference Time
import time
model = SSAD(cfg).cuda()
data = torch.randn((1, 303, 512)).cuda()
t1 = time.time()
for i in range(1000):
res = model(data)
t2 = time.time()
print("Inference Time:", (t2 - t1) / 1000) # 0.00236631178855896